MLDashboard

Monitoring solution for tensorflow training. Particulary useful for image classification models. Not compatible with google colab or other notebook based runtimes.

Getting Started

This guide assumes you already understand python and tensorflow.

Installation

pip install MLDashboard

Examples

See Examples for usage.

Quick Start

To start, you need a dashboard.json config file. This should be in the same directory as your script. Here is an example:

{
    "modules":[
        [
            ["LossMetricsGraph", {}],
            ["LossMetricsNumerical", {}]
        ],
        [
            ["StatusModule",{}],
            ["EmptyModule", {}]
        ]
    ]
}

NOTE: All code in this demo should be protected by

if __name__ == '__main__':

to prevent multiprocessing conflicts.

The dashboard can easily by added to an existing machine learning project. Import the dashboard as shown.

from MLDashboard.MLDashboardBackend import createDashboard
from MLDashboard.MLCallbacksBackend import DashboardCallbacks, CallbackConfig
from MLDashboard.MLCommunicationBackend import Message, MessageMode

Before training starts, create the dashboard.

#MAKE SURE YOU HAVE A DASHBOARD.JSON FILE IN THE SAME DIRECTORY AS YOUR SCRIPT
dashboardProcess, updatelist, returnlist = createDashboard(config='dashboard.json')

Connect the callbacks to your training.

config = CallbackConfig()
labels = list(range(0,10)) #labels should be customized for the data. This is for mnist number recognition
callback = DashboardCallbacks(updatelist, returnlist, model, x_train, y_train, x_test, y_test, labels, config)

model.fit(x_train, y_train, epochs=10, callbacks=[callback])

After training ends, you can send evaluation stats to the dashboard.

model.evaluate(x_test, y_test, batch_size=128, callbacks=[callback])

To exit the dashboard cleanly, use the following code:

updatelist.append(Message(MessageMode.End, {}))
print("Exiting cleanly...")
dashboardProcess.join()
print("Dashboard exited.")
#This handles any extra data that the dashboard sent, such as save commands
callback.HandleRemaingCommands()

Here is a full example with python code:

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = '2' #stops agressive error message printing
import tensorflow as tf
from tensorflow import keras
from MLDashboard.MLDashboardBackend import createDashboard
from MLDashboard.MLCallbacksBackend import DashboardCallbacks, CallbackConfig
from MLDashboard.MLCommunicationBackend import Message, MessageMode

def run():
    print("Starting interactive dashboard demo...")
    print("Setting up dashboard...")

    #Create dashboard and return communication tools (this starts the process)
    #MAKE SURE YOU HAVE A DASHBOARD.JSON FILE IN THE SAME DIRECTORY AS YOUR SCRIPT
    dashboardProcess, updatelist, returnlist = createDashboard(config='dashboard.json')

    print("Loading data...")
    (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()

    print("Formatting data...")
    x_train = x_train.reshape(-1, 784).astype("float32") / 255.0
    x_test = x_test.reshape(-1, 784).astype("float32") / 255.0

    print("Sampling data...")
    # Limit the train data to 10000 samples
    x_train = x_train[:10000]
    y_train = y_train[:10000]
    # Limit test data to 1000 samples
    x_test = x_test[:1000]
    y_test = y_test[:1000]

    print("Creating model...")
    model = keras.Sequential([keras.layers.Dense(128, activation='relu'), keras.layers.Dense(10)])

    model.compile(optimizer='adam', metrics=["accuracy"], 
                  loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True))


    print("Creating callbacks...")
    #Callbacks require update and return list for communicating with dashboard
    #Model and datasets are useful for sending that data to certain modules
    config = CallbackConfig()
    labels = list(range(0,10))
    callback = DashboardCallbacks(updatelist, returnlist, model, x_train, y_train, x_test, y_test, labels, config)

    model.fit(x_train, y_train, epochs=50, callbacks=[callback])

    print("Evaluating model...")
    #This is connected to the callback so the data is sent to the dashboard
    model.evaluate(x_test, y_test, batch_size=128, callbacks=[callback])

    updatelist.append(Message(MessageMode.End, {}))
    print("Exiting cleanly...")
    dashboardProcess.join()
    print("Dashboard exited.")
    #This handles any extra data that the dashboard sent, such as save commands
    callback.HandleRemaingCommands()

if __name__ == '__main__':
    run()

Other guides:

• Customizing the Dashboard
• Creating a Custom Module
• Creating Custom Callbacks (advanced)
• Module Documentation Coming soon!
• Data Structure Documentation
• Primary Functions Documentation